A vendor selling signal quality opened with eight words about my company that were false. The signal was real: a post I wrote ten months earlier about another company’s award. The pipeline took the subject of the document and handed it to the author, then generated a dated membership history on top of the error, with nothing in between to catch it.
The commentary literature on automated sales development is close to worthless as evidence. It is produced by companies with a position, it quantifies nothing that can be checked, and it stays at the level of tone and reply rates. What almost nobody publishes is a single failure traced all the way back: the message, the vendor’s own explanation of the mechanism, and the source document that actually produced it. That is what this report is. On 12 August 2026 a seller at WeTransact, a French firm that helps software vendors publish on the Microsoft commercial marketplace, opened a message by congratulating me on a Microsoft for Startups membership my company did not have. When I asked where the claim came from, he named the signal type, and with that one sentence I found the source in about ten minutes. It was a post I had written ten months earlier about IgniteData being selected for the Pegasus Program. The system read the document correctly, identified the right program, and then attached it to the wrong entity, choosing the author over the subject. Every founder who writes about their own industry is exposed to that error, which is most of them.
Ten findings, each labeled by evidence type and carrying a numbered source. The company is legitimate and the disconfirming findings are left in, because a report that only points one way is not a report.
WeTransact was founded in 2023 by former Microsoft staff, is based in Lyon, sells a Microsoft marketplace platform, appears in Microsoft’s own published customer stories, and was named Microsoft Startup of the Year for France in 2025. Anyone reading this as a scam warning has misread it.
The opener stated that BeaconSpec had recently joined Microsoft for Startups and that I was “a few months in” with credits as “step one.” The Azure portal that morning said: “This account can’t be used for Microsoft for Startups credits.” No balance, no application banner, no member state of any kind. Both screenshots are reproduced rather than described.
Asked where the claim came from, the vendor said it came from LinkedIn posts by someone at my company mentioning the program. It did. Ten months earlier I posted commentary on IgniteData being selected for the Microsoft for Startups Pegasus Program. The system took the subject of the post and assigned it to the author.
A post about another company’s award cannot yield “recently,” “you’re a few months in,” or “the credits were step one.” None of that was in the source. The retrieval step misidentified the subject and the generation step invented a membership history around the error, with no check between them.
The homepage offers “AI messaging and signals to create new opportunities” and account data on “who has budget, and who controls it.” Microsoft’s customer story describes the platform as built on Azure OpenAI Service. A firm selling signal quality sent a message whose signal was wrong.
Microsoft’s December 2024 story cites over 300 ISVs. The About page says 450+. The homepage says 500+. The last two were live on the same site on the same day I checked. None is audited and none states what counts as an accelerated ISV.
LinkedIn’s Professional Community Policies say “Do not share content that is false, misleading, or intended to deceive.” EU law has banned misleading business-to-business advertising since 2007, and the definition of advertising is wide enough to reach a sales message. Neither produces a consequence at this unit size.
Article 50 of the EU AI Act became applicable on 2 August 2026, ten days before the message was sent. Its text-labeling duty is aimed at content published to inform the public on matters of public interest, and its marking duty falls on model providers. A private sales message sits outside both.
The proposition is data an ISV cannot verify alone: which accounts hold committed Microsoft spend, who controls the budget, which sellers to call. A buyer gets exactly one free test of that data discipline, and it is the claim the vendor makes about the buyer’s own company. That test was failed, publicly, in writing, in the first sentence.
He replied the next morning naming the mechanism accurately without being able to verify it, offered a same-named company as the likely mix-up, and promised a signal review. There was no apology. The homonym theory is refuted by the post itself: it was not another BeaconSpec, it was another company entirely, named in my own text.
This is the whole finding in one place. The left side is a document anyone can read. The right side is what arrived in my inbox. Each row is a separate failure, and only the first one is a retrieval problem.
The post says IgniteData was selected for the Microsoft for Startups Pegasus Program. My company appears in it once, as my own job title in the byline. The system read the document correctly and then handed the program to the person who wrote about it. There is no name collision to blame and no bought database to blame.
Gap: the entity in the output is not the entity in the source
The post is dated ten months before the message. Nothing in it supports recency, and the real date contradicts both phrasings. This is not a retrieval error carried forward. It is a duration manufactured to make the opener feel researched.
Gap: generated specifics with no source record behind them
The post mentions no credits, no spend, and no benefits of any kind. A system that had merely misattributed would have produced a vague and roughly harmless sentence. This one produced a specific, dated, benefit-bearing account of a membership that does not exist.
Gap: the difference between a bad lookup and a fabrication
Pegasus is a specific selective track. The message flattened it into ordinary program membership, then extrapolated a graduation the sender could plan around. Each step is individually plausible, and the compound result is a paragraph of confident detail about a company state that never existed.
Gap: plausibility compounds faster than accuracy
The message is reproduced complete in the report, including the repetition, so nothing rests on my characterization of it. It asks “Worth 15 min?” twice in identical words, once mid-paragraph and once at the end. A human writing 78 words does not ask the same closing question twice, and two closers in one block is the signature of a template assembled from parts where no pass removed the duplicate. Treat a duplicated call to action as a canary: when a template emits the same ask twice, no human read the output, and if no human read the output then no human checked the premise. The message also contains no question about me at all. Every sentence is a declarative statement about my company, my timeline, my priorities, or what I should do next. The only interrogative in the whole thing is the request for a meeting, asked twice. That is the tonal signature of assumed authority, and it is what makes the error land as an insult rather than a typo.
The report ends where a case study should end, on things a reader can act on this week without hiring anyone.
Most cold outreach contains exactly one falsifiable assertion about you, and testing it takes less time than reading the rest of the message. If it fails, you have learned something about the sender’s data discipline that no case study would have told you, and you learned it for free. If you are evaluating any vendor whose product is data you cannot audit, pick three signals on accounts you know well and ask them to show where each came from and when it was last refreshed.
Not a confidence score, a record. Facts about the recipient come from a retrieval step with a record identifier attached, the generator is permitted to phrase them and forbidden to originate them, and any message containing an unsourced factual assertion fails a check before send rather than after. That gate is cheap to build and it is the only part of the stack that protects the brand.
“Are you in Microsoft for Startups? If so, this is the moment to set up the marketplace side” carries the same hook, works on the same segment, and cannot be false. The sender gives up a small amount of warmth and eliminates the entire failure mode. That trade is available to every team running this play and almost nobody takes it.
This isn't a vendor summary. Every sentence is labeled by what stands behind it: verified fact, vendor claim, third-party estimate, my assessment, hypothesis, or scenario. Sources are numbered and clickable. Forward-looking sections use scenarios with observable tripwires, not forecasts. It's the same method behind every market assessment I write.
Twenty-one pages, built from public sources with no client brief and no interviews. Read it in the browser or take the PDF.
Each report here answers a real question, directed and researched against public sources and evaluated against a stated assumption, then delivered as Word and PDF. If you're weighing a platform, sizing a category, or defending a number to a board, tell me the decision behind it and I'll tell you honestly whether a report is the right tool.
Commission an assessment